model: "RetinaUNetC008"
trainer: "DetectionTrainerPolyLR"
predictor: "BoxPredictorSelective"
plan: D3C002_3d

planners:
  2d: [D2C002]
  3d: [D3C002, D2C002]

augment_cfg:
  oversample_foreground_percent: 0.5 # ratio of fg and bg in batches
  augmentation: ${augmentation}
  dataloader: "DataLoader{}DFast"
  dataloader_kwargs: {}

trainer_cfg:
  # Per default training is deterministic, non-deterministic allows
  # cudnn.benchmark which can give up to 20% performance. Set this to false
  # to perform non-deterministic training
  deterministic: True
  fp16: True # enable fp16 training. Makes sense for supported hardware only!
  eval_score_key: "mAP_IoU_0.10_0.50_0.05_MaxDet_100" # metric to optimize

  num_batches_per_epoch: 2500 # number of train batches per epoch
  num_val_batches_per_epoch: 100 # number of val batches per epoch
  max_num_epochs: 50 # max number of epochs # CHANGE TEMP
  overwrites: {}

  initial_lr: 3.e-4 # initial learning rate to start with
  weight_decay: 3.e-5 # weight decay for optimizer

  warmup: 4000 # number of iterations with warmup
  warmup_lr: 1.e-6 # learning rate to start warmup from

model_cfg:
  matching:
    # IoU Matcher Parameters
    fg_iou_thresh: 0.4 # IoU threshold for anchors to be matched positive
    bg_iou_thresh: 0.3 # IoU threshold for anchors to be matched negative
    # If ground truth has no matched anchors, use the best anchor which was found
    allow_low_quality_matches: True

    # ATSS matching
    num_candidates: 4
    center_in_gt: False

  hnm: # parameters for hard negative mining
    batch_size_per_image: 32 # number of anchors sampled per image
    positive_fraction: 0.33 # defines ratio between positive and negative anchors
    # hard negatives are sampled from a pool of size:
    # batch_size_per_image * (1 - positive_fraction) * pool_size
    pool_size: 20
    min_neg: 1 # minimum number of negative anchors sampled per image

  plan_arch_overwrites: # overwrite arguments of architecture
      strides: [[1, 2, 2], [1, 2, 2], [2, 2, 2], [2, 2, 2], [1, 2, 2]]
      conv_kernels: [[1, 3, 3], [1, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3], [1, 3, 3]]
      decoder_levels: [2, 3, 4, 5]

  plan_anchors_overwrites: # overwrite arguments of anchors
    width: [[2.0, 3.0, 4.0], [4.0, 6.0, 8.0], [8.0, 12.0, 16.0], [8.0, 12.0, 16.0]]
    height: [[3.0, 4.0, 5.0], [6.0, 8.0, 10.0], [12.0, 16.0, 20.0], [24.0, 32.0, 40.0]]
    depth: [[3.0, 4.0, 5.0], [6.0, 8.0, 10.0], [12.0, 16.0, 20.0], [24.0, 32.0, 40.0]]